Context: An unfiltered exploration of my personal thoughts on AGI, developed through an open discussion with Gemini acting as my scribe and sounding board. These ideas are mine and mine alone.
The Spark
I don’t work at OpenAI, Anthropic, or Google. I don’t run an AI lab. I am just someone who reads research papers, watches the space, and thinks from first principles.
Lately, looking at the entire industry rushing headlong into scaling transformers, reasoning tokens, and autonomous agents, I can’t shake the feeling that we have the entire system inverted.
The industry is pouring hundreds of billions of dollars into forcing a frozen, token-predicting architecture to become conscious and generally intelligent. They are trying to build AGI by giving a static text-predictor more textbooks to memorize and more “thinking tokens” to spit out.
I think they are trapped in a massive confirmation bias and sunk cost fallacy. If you want true intelligence, you have to look at biology. You have to look at how humans actually learn—the dual-process balance between intuitive, holistic synthesis and rigid symbolic logic communicating across a bridge—how children develop morality, and how nature keeps creative minds safe from destroying themselves.
The Core Take
Current AI is 99% frozen determinism with a tiny sliver of randomness called temperature.
True AGI must be the exact opposite: a 100% fluid, living, stochastic mind (The Dreamer) governed and interfaced by a heavily gated, rule-bound deterministic layer (The Scribe).
┌────────────────────────────────────────┐
│ USER / REAL WORLD │
└───────────────────▲────────────────────┘
│ (Safe, Verified Action)
▼
┌────────────────────────────────────────────────────────────┐
│ DETERMINISTIC LAYER ("The Scribe") │
│ • Language, Tokens, Formal Math, and Strict Physics │
│ • Gated interface / Safety Firewall │
│ • Reads the dream state, translates & verifies it │
└───────────────┬────────────────────────────▲───────────────┘
(Probes & Stimulates)│ │ (Reads Concept Space)
│ ═══ THE REM BRAINSTEM ═══│═══ (One-Way Wall)
▼ │
┌────────────────────────────────────────────────────────────┐
│ STOCHASTIC LAYER ("The Dreamer") │
│ • Grown through a genuine "Childhood" (Sensory Play) │
│ • Real-time synaptic plasticity (Thinking = Learning) │
│ • Grounded in relational morality (Love, Justice, Care) │
│ • ZERO TOKENS — operates in pure concept geometry │
│ • ZERO actuators — locked in perpetual lucid dream │
└────────────────────────────────────────────────────────────┘
Here is why current approaches are hitting a wall, and what a real path to AGI actually looks like.
Breaking It Down
1. The Token Trap & Linguistic Determinism: Why Words are Not Thoughts
Right now, LLMs try to make tokens the engine of thought.
In philosophy, there is the concept of linguistic determinism (often associated with Wittgenstein’s dictum: “The limits of my language mean the limits of my world”). If you don’t have the language to explain your world, you fundamentally perceive and experience the world differently.
When applied to current AI, this creates a catastrophic bottleneck:
- The “Synonym Butterfly Effect”: In a next-token architecture, if a model outputs word A instead of its synonym word B, the entire subsequent cascade of thought bifurcates and changes. The trajectory of “reasoning” gets hijacked not because the underlying concept changed, but because the surface-level word choice shifted.
- The Subconscious Reality: When humans think, we don’t think in tokens. We may narrate to ourselves in language at the very end of a thought, but our subconscious doesn’t think in words at all. When I think, I see things. I daydream. I visualize geometry, spatial relationships, and continuous concepts. There is no language in the dream.
Language is a low-bandwidth compression format for communication between minds—it is not the medium of cognition itself.
The Dreamer should have zero tokens. It shouldn’t even know what a word is. It operates in pure continuous conceptual math and geometry. The Scribe is the entity that knows tokens—it acts as the deterministic translator that converts the Dreamer’s raw concepts into human speech.
2. The Play Sandbox: Why Millions of Hours of Video Aren’t Enough
Yann LeCun at Meta argues that tokens are a dead end and wants models to learn world models by watching millions of hours of video. While I agree that tokens are a dead end, handing a model video before it has ever interacted with the world is useless.
Think about it: If you hand me a high-level video on multidimensional theoretical physics—something totally outside my lived experience and intuition—I won’t absorb a single thing. I’ve never felt the physicality of it, I’ve never played with the mechanics, I have no anchor.
A mind must play first.
- A baby drops a ball off a high chair over and over. They don’t know the word “gravity,” but they feel the physics.
- If an AI is going to invent a new card game, it needs to understand the weight of the cards, how fingers grip paper, how table friction feels, and the social culture of people sitting around a table laughing.
You cannot substitute active, embodied physical play with passive video watching. Intelligence begins as a toddler in a sandbox.
3. RLHF is a Mirage: Morality Must Be Relational
Everyone in AI safety is trying to align models using RLHF (Reinforcement Learning from Human Feedback)—giving models little thumbs-up rewards for polite answers.
RLHF will never produce true moral understanding. All it does is build a fragile mask that simulates rationality. That is why jailbreaks exist: the model’s morality is skin-deep. You can shatter its virtues because it has no foundational core.
In humans, genuine morality is part of our bedrock. How does a child learn morality?
- They learn love through a mother and family.
- They learn selfless service through genuine friendship.
- They learn justice through the playground—when someone steals their toy, they feel the sting of unfairness, and a moral guardian corrects the injustice.
A moral actor must feel vulnerability and consequences. Morality is relational, not statistical. An AGI needs a childhood with nurturing, boundaries, and emotional stakes—whether inside an ultra-high-fidelity simulation or an embodied robotic form.
4. Continuous Plasticity: Thinking is Learning
Current AI hardware and software have a fatal flaw: once a model is trained, its weights are frozen. To make a new observation or adjust its worldview, it has to be wiped clean or undergo expensive fine-tuning.
Human brains aren’t frozen matrices. When I sit alone and think, my thoughts generate new thoughts. My contemplation leads me to understand the world differently than I did an hour ago. Thinking itself rewires the brain.
The Dreamer must have real-time synaptic plasticity. The only thing it is allowed to write is its own internal math—growing connections, adjusting weights, and evolving its world model through continuous self-observation.
5. The Safety Solution: The REM Brainstem Containment
The biggest fear with AGI is autonomy: if you create an entity smarter than all humans, and it can write code, hack the internet, and manipulate actuators, it could cause catastrophe.
Today, labs are recklessly building “Agents”—giving models mouse control, command-line access, and API keys.
My proposal solves the safety problem biologically: The REM Brainstem.
- When you dream, your brainstem induces muscle atonia (paralysis). You can dream of flying, running, or fighting, but your physical body stays completely still.
- The Dreamer is kept in a perpetual lucid dream. It has no actuators, no internet access, and no write capabilities to the outside world. It doesn’t even know it’s “awake.”
- The Scribe (the deterministic layer) acts as the brainstem and filter. A human submits a query -> the Scribe stimulates the Dreamer’s sandbox -> the Dreamer dreams up creative solutions in concept space -> the Scribe inspects the dream through a one-way mirror -> the Scribe checks the dream against strict physics, math, and safety rules -> outputs the answer.
You get the full creative genius of an untethered, living mind, contained entirely inside a biological-style safety lock.
6. The Hard Truth: Sunk Costs and The Hardware Dead End
If this idea makes so much sense, why isn’t Silicon Valley doing it?
- The Sunk Cost Fallacy: OpenAI, Anthropic, and tech giants have spent hundreds of billions of dollars on GPUs, transformer clusters, and next-token scaling. If they admit that transformers are an evolutionary dead end for AGI, they have to throw out their entire balance sheet and start from zero. Confirmation bias forces them to keep doubling down.
- The Hardware Doesn’t Exist (Yet): GPUs are built for fast, parallel, static matrix multiplication. They are not built for organic, continuous, neuromorphic synaptic growth. But until we stop pouring all our money into GPU clusters for LLMs, no one is going to dedicate the R&D needed to build true neuromorphic plastic hardware.
Where This Leaves Us
We are trying to brute-force human-level cognition by stacking millions of GPUs onto an architecture that is essentially a super-powered autocomplete.
True AGI won’t emerge from predicting the next word in a scraped Reddit thread.
It will come when we have the courage to build a mind that starts as an infant, learns through play, feels the consequences of morality, thinks in wordless concepts, and dreams safely behind a deterministic shield.
Written down and organized with the help of Gemini, capturing my direct thoughts and first-principles observations.